Gradient Descent in Forward Pass

DoubleAI introduces "mechanism intervention," placing a coding agent inside training and inference so it can inject Python code into the model backbone to take over computation instead of only tuning weights. With roughly 20 lines injected, a pretrained model performs gradient descent within a single forward pass; stacking 10 code layers lifts accuracy on 30 classic algorithms from 55% to 84%. Separately, a 763B-parameter model reads its own deep activations, detects drifting reasoning and rolls back to recompute, gaining about 7 points on AIME and HMMT 2026.